Status of the Embedded GPU Space in Linux Robert Foss Collabora Ltd. [email protected] Abstract—Embedded graphic processing units (GPUs) are support? Or if the support comes at an unacceptable used for increasingly diverse types of applications, placing new cost? Or if the vendor exits the graphics business? requirements on the GPUs, only some which are supported by vendors. While most IP vendors still don't support their devices Many drivers have come as a result of reverse with Open Source drivers, in many case the Open Source community has stepped up to provide these. engineering (RE) efforts by the Open Source community and are now maintained as a part of the In this article the current landscape of embedded GPUs and Linux kernel, Mesa, and libdrm. Often the current challenges with an emphasis on their respective Open motivation behind the RE effort stems from a Source support will be discussed. desire to fully enable powerful hardware that is Keywords—virtual reality; vulkan; open source; linux; already available, for example in cheap graphics; driver; development kits or single board computers (SBCs). I. INTRODUCTION New standards like Vulkan also have a large The embedded GPU space is seeing rapid impact on the software stack, with some changes with the introduction of new standards like implementations going as far as implementing Vulkan[1] and new applications like VR and legacy standards like OpenGL on top the Vulkan wearables. stack. This, of course, affects the GPU hardware and This paper will look at different GPUs and what we expect from it, which in turn means that provide an overview for their Linux support, the software ecosystem is seeing many new including a comparison of drivers provided by the demands placed on it that spans from low latency vendor against the one provided by the Linux output [2] of assisted reality and virtual reality (XR) Community, looking at the benefits and to low power requirements of wearables [3]. disadvantages of each of them. The motivation for product manufacturers to choose using a fully Open Source stack comes II. NEW DEVELOPMENTAL TARGETS down to the fact that in most cases the proprietary stack doesn't meet their needs, including stability, A. Virtual Reality performance and features. Assisted reality and virtual reality (XR) is a new For some manufacturers other factors come into type of frontier for graphics stacks and input stacks. play, like the ability to provide decade long support The importance of latency has previously been for their products. In order to provide a secure much lower, and as a result both graphics drivers operating system, the graphics stack has to be kept and entire graphics stacks are not optimized for this up to date. For this to happen either an Open Source use case. XR also places similar requirements on stack has to be used, or the GPU-vendor has to be the input stack, where latency can become an issue contracted to provide support for the term of your too. Unlike most features low latency puts system- product support. To product manufacturers this wide requirements on the graphics stack, to which introduces a risk and a financial burden. What there are no simple fixes. happens if your vendor chooses to no longer offer www.embedded-world.eu A partial solution, called DRM Leasing[4], has been put forward by Keith Packard. It allows applications to take ownership of hardware components directly, thereby bypassing traditional components of the graphics stack like window servers and desktop environments. By doing so many latency introducing components can simply be avoided. This approach is based on flagging certain display components in the kernel as unavailable for normal applications and then making these components available directly to user space Open Source Linux drivers have existed for a long time, especially for Desktop class GPUs, but applications. III. during the past few years supportGPU Dfor Embedded class GPUs has materialized too. Other components like GPU shader-compilers, human interface devices (HID) and displays RIVERS themselves all have to be optimized for low-latency in order to provide the best possible user 4 0 Intel experience. 0 2 Embedded Graphics Drivers B. Vulkan Figure 1: Timeline of Open SourceAMD driver becoming available for vendors. NVidia Vulkan offers a new tier of low-level GPU access, and with it a lot of flexibility. This brings Qualcomm Adreno xxx In Fig. 1 the date of introduction for the driversA with it the possibility of being more power efficient of GPU vendors has been listed. A closerBro ainspectiondcom VCx if the new API is utilized correctly on hardware that of the second half reveals that most to EmbeddedVivante GCxx ARM Mali-4xx supports it. For example power savings of up to GPUs now have a viable Open Source drivers. The ARM Mali-7/8xx 20% can be achieved for multimedia use-cases[5]. latest additions, the ARM T7/T8/Gxx drivers, are 0 2 0 rather new, but quite close to being included in the 2 Vulkan also enables novel use-cases like implementing OpenGL on top of Vulkan in order to normal Linux graphics stack. always have an OpenGL driver for any Vulkan 9 0 0 enabled GPU, like, e.g. Zink[6], which may be 2 desirable to lessen the burden on maintaining an Broadcom OpenGL driver when a vulkan driver is already 9 0 0 available. 2 VC4 Qualcomm Industry adoption of Vulkan is keeping a high 9 Driver added pace. Android[7] is moving over to using Vulkan as 0 A200 0 2 A300 VC5 its main graphics API, which means not only Vivante A400 V3D making it available to applications, but also moving 9 1 0 2 Initial RE A 0 all internal Android 3D-rendering over to Vulkan. 9 5 0 0 0 2 ARM A600 Figure 2: Timeline of embedded GPU driver development From the standpoint of a product manufacturer Initial RE Mali-2/4xx Driver added 9 1 Vulkan offers some new benefits, like the GC2000 0 Conformance Test Suite (CTS) being available as an GC3000 2 As can be seen from Fig. 2, driver support for GC7000 Open Source project[8]. This enables manufacturers different vendors vary, only the ones that target the 9 1 to ship product and have high levels of confidence 0 embedded space and offer some Open Source 2 in the stability and standards conformance of their support have been listed. Currently, the full rangeIn ioftial RE MaliT7/T8/Gxx products. Broadcom GPUs, Qualcomm GPUs and Vivante Driver added 9 GPUs are supported, while support for ARM GPUs 1 0 is still in the early stages. 2 Another topic related to these drivers is how long drivers stay supported. Naturally that differs between GPUs, but for example the AMD R100 series of GPUs were launched in 2000[9], and the Open Source driver is still maintained and receives patches[10]. Notable exceptions from this list are NVidia and Imagination Technologies GPUs. While both of them make GPUs targeting the embedded space neither have participated in supporting Open Source drivers, and NVidia actively prevents the driver that support for their newest GPUs which are still only exists from being usable by contractually limiting sparsely available in the marketplace today. the required firmware binaries from being Due to the similarities between different series of distributed with the Open Source driver. It should Vivante GC GPUs, the effort required for be noted however that this does not apply to some supporting a new GPU has been relatively low NVidia Tegra GPUs, where firmware is freely historically[14]. distributed. With this overview out of the way, we can have a look into the status of the individual drivers. glmark2 score A. Broadcom Etnaviv GC3000 The VC4 and V3D[11] drivers were created by Eric Anholt and are being actively maintained by him and supported by Broadcom. Vivante GC3000 The initial VC4 driver development was started as a result of it being the GPU shipped in the 0 100 200 300 400 500 600 700 Raspberry Pi series of SBCs. Since then the VC5 & Higher is better VC6 GPUs have had support added in a driver called V3D. Figure 3: Vivante & Etnaviv GC3000 benchmark Currently the driver is very mature, and it offers The performance of the Open Source driver (Fig. the best GPU-vendor support out of the embedded 3) is comparable to the proprietary one, and has drivers. been improving much over time. B. Qualcomm This driver has to this moment not received any The Adreno driver, freedreno[12], originated support from its hardware designer. from the Open Source ecosystem and Rob Clark’s efforts, and is by now a very mature driver. D. ARM Additionally, Qualcomm is one of the very few The ARM drivers, lima[14] and panfrost[15], embedded hardware vendors that actually support have been a long time coming. Initial work on the their Open Source driver to some extent. This reverse engineering of the Mali-2/3/4xx series of support could be more extensive, but it is still far GPUs was presented at FOSDEM 2012 by Luc ahead of the competition in the embedded space Verhaagen[16], but the work was never completed listed below. or submitted to the Linux graphics stack projects. Nevertheless, recently it has been restarted and it is Support for their hardware is complete, in that it now close to the point were it will be included in is offered for the complete A2-6xx range of GPUs the Linux graphics stack. (Fig. 2). Open Source driver support for the Mali-T7/T8/ As Qualcomm is an SOC manufacturer, the Gxx series has come a very long way in the last two Adreno GPUs are only available in their years.